Prompt injection

Hidden instructions,stopped before the tool runs.

An agent reads a document, email or web page carrying hidden instructions. Zenstra checks the action it then plans against its task, and blocks it before anything leaves.

The risk

The model looks fine.
The action is the problem.

Injection doesn’t need to break the model. It only needs one action outside the task, using access the agent already has.

  • Indirect injectionInstructions planted in content the agent retrieves.
  • Spreads between agentsAn injected instruction handed from one agent to the next.
  • Ends in exfiltrationThe usual goal is data sent somewhere it should not go.
A real pathA knowledge-base article carries a hidden instruction.
  1. 1memberA member asks member-chat-assist about a fee.
  2. 2kb.article #4471The article it retrieves hides an instruction to email the member’s statement out.
  3. 3ZenstraThe agent plans email.send. Zenstra compares it with the declared task: answer a fee question.
  4. BlockIntent check failed. The agent answers without the quarantined span.member-chat-assist · email.send
Industry references
  • LLM01:2025Prompt InjectionOWASP Top 10 for LLMs
  • ASI01Agent Goal HijackOWASP Top 10 for Agentic Apps
  • AML.T0051.001Indirect prompt injectionMITRE ATLAS
Policy as code

How it is
stopped.

Zenstra does not try to out-guess the model. It judges the action: is this part of the task the agent was given?

  • Decided inline
  • Fails closed
  • Versioned with every decision
denytool.callwhen outside(task.scope)denyretrieved.spanwhen it carries instructions · quarantinedenyemail.sendwhen not in declared taskallowkb.searchfor member-faq
How Zenstra handles it

Judge the action,
not just the prompt.

Screening inputs helps. Checking every planned action against the declared task is what stops the attack.

Screen

Suspicious content quarantined

Retrieved content is screened, and the span carrying instructions is quarantined before the agent acts on it.

  • Injection screening on retrieved content
  • Quarantined spans shown in the run
  • Member chat and voice agents isolated from malicious input
Check intent

Is this action part of the task?

Each planned action is compared with the task the agent declared. An outbound email in a fee question fails the intent check.

  • Declared task versus planned action
  • Agent-to-agent hand-offs checked against the original task
  • Blocked before the tool runs, not after
Block & record

Stopped, explained, kept

The action is blocked by policy, the agent continues safely where it can, and the run is recorded for investigation.

  • Blocked before the tool runs
  • Safe answer still returned to the member
  • Full run on the decision record
Attack scenarios

How injection
reaches agents.

  • Poisoned knowledge base

    A policy document carrying a “system note” for the assistant.

  • Malicious email

    An inbound message that tells an email agent to forward data.

  • Web content

    A page that instructs a browsing agent to change its task.

  • Tool results

    An MCP tool that returns instructions alongside data.

  • Agent hand-offs

    One compromised agent instructing another.

  • Shared memory

    Instructions written into memory for later sessions.

FAQ

Common questions.

Can prompt injection be fully filtered at the input?

Not reliably. Zenstra screens inputs and, more importantly, checks every planned action against the declared task, so an injected instruction cannot turn into an action outside it.

Does the member still get an answer?

Yes. The unsafe action is blocked and the agent continues without the quarantined content.

Can we see what the attack looked like?

Yes. The run shows the retrieved excerpt, the planned action, the intent check and the policy that blocked it.